How to Choose an AI Development Company: 12 Questions to Ask Before You Sign

The short answer

Pick the AI development company that can show you a system running in production with real users, names the engineers who will write your code, explains how it will test the AI’s answers before launch, and puts in writing that you own everything it builds. Budget $15,000 to $40,000 for a focused AI app or agent, expect it live in four to eight weeks, then allow $200 to $1,500 a month to run it.

An AI development company is a team that designs, builds and maintains software in which a language model or other machine learning model does part of the work, connected to your own data and systems. Plenty of agencies now use that label. The twelve questions below are the ones we’d ask if we were on your side of the table, and they work just as well on us.

Do you need an AI development company at all?

Sometimes the honest answer is no. Before you shortlist anyone, check whether one of these fits your situation better.

• A freelancer, if the job is configuring an existing platform or a small prototype with one integration. It costs less, but one person is a single point of failure when something breaks
• An in-house hire, if AI is your product rather than a feature and you’ll keep improving it for years. In the Bay Area that means a senior salary and a long search, so it only pays off with a multi-year roadmap
• A large consultancy, if you need change management across thousands of staff, a formal audit trail for regulators, or a vendor your procurement team already approves. You’ll pay for that extra layer
• Software you already own, if the job lives inside one tool. Many CRMs and help desks now ship AI features, and a month of testing them costs almost nothing

Questions 1 to 3: is the work real, and who does it?

Agencies fit the middle ground: a defined business problem, several systems to connect, and no appetite to build a team. Our build vs buy guide covers that decision in more depth. Once you’ve settled on an agency, start with proof.

Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, and estimated that only about 130 of the thousands of vendors selling “agentic AI” actually deliver it. Demos are cheap. Production is where vendors separate.

• 1. Can you show me a system in production with real users, and can I speak to that client? A demo runs on clean sample data. A live system has logs, odd edge cases and a client who can tell you what broke in month two
• 2. Who will write the code, and can I meet them before I sign? Get names and roles in writing. The person on the sales call often isn’t the person in the repository
• 3. Will any part of the build be subcontracted? Offshore engineers aren’t a problem in themselves. Not knowing who can see your data is

Questions 4 and 5: how will they test the AI?

Most buyer guides skip this, yet it predicts the outcome better than any portfolio. AI output is probabilistic. The same question can get a slightly different answer tomorrow, so “it worked in the demo” proves very little.

Two questions do the work here. If the answer to the first is “we test it thoroughly”, keep pressing. Teams that test properly can open a results sheet from a past project in a couple of minutes.

• 4. How will you prove the system works before launch? The answer you want is an evaluation set: thirty to a few hundred real examples from your business, each with the correct outcome, rerun every time the prompt, model or data changes. Ask what pass rate counts as ready and who signs it off
• 5. What happens when the AI is wrong? Listen for confidence thresholds, a handoff to a person, an approval step before anything is sent or paid, a log of every decision, and a way to roll back

Questions 6 to 8: where does your data go?

Every AI system sends some of your information to a model provider, so you need to know which one, on what terms, and who else can see it.

• 6. Which model providers will see our data, and on what retention terms? OpenAI’s API documentation, for example, says data sent to the API is not used to train its models unless you opt in, and that abuse monitoring logs are kept for up to 30 days by default. Zero Data Retention exists, but only for eligible customers OpenAI approves. Ask for the equivalent terms of every provider in the design
• 7. Do you have a SOC 2 report, and who holds the API keys? SOC 2 is an examination by an independent CPA, under AICPA standards, of controls over security, availability, processing integrity, confidentiality or privacy. Few small agencies have their own report. That’s fine if the system runs in your cloud accounts on providers that do, with keys in your company’s name
• 8. Which privacy laws apply, and will this system make significant decisions about people? In California, that question now has a deadline attached

Which US privacy laws touch an AI project?

The California Consumer Privacy Act, as amended by the CPRA, covers for-profit businesses that meet one of several thresholds. One of them is annual gross revenue above $26,625,000, the figure the California Privacy Protection Agency set from January 1, 2025.

The newer rules matter more for AI. Regulations on automated decisionmaking technology (ADMT), risk assessments and cybersecurity audits were approved by California’s Office of Administrative Law on September 22, 2025. Under section 7200, a business using ADMT for a significant decision about a consumer before January 1, 2027 must comply with the ADMT article by that date. A significant decision is one that grants or denies financial or lending services, housing, education, employment or contracting work and pay, or healthcare services.

In practice, an AI that screens job applicants, scores loan requests or triages patients is in scope and needs a pre-use notice plus opt-out and access rights. A chatbot answering delivery questions usually isn’t. Other states have passed their own privacy laws and some are drafting AI rules, so ask which ones the agency designed for, then confirm with your lawyer. Your agency isn’t a law firm, and the good ones say so.

Questions 9 and 10: who owns the work, and how is it priced?

9. Who owns the code, prompts, evaluation sets and accounts? The contract should assign all of it to you on payment, including any fine-tuned model weights, with cloud, model provider and repository accounts registered to your company from day one. One wrinkle: the US Copyright Office’s January 2025 report on copyrightability concluded that purely AI-generated material isn’t protected by copyright, and that prompts alone don’t give a person enough control to count as the author. Own the code base, and protect what makes the product valuable through contracts and trade secrets too.

10. Is this fixed price or time and materials? Under time and materials you pay for hours worked, so you carry the risk if the estimate was wrong. Under fixed price the agency carries it, which only works with a written scope and a change-request process. Open-ended research suits time and materials. On a defined business system, fixed price keeps both sides honest, and it’s how we quote every Zeebrix project after a free discovery call.

Questions 11 and 12: what will it cost, before and after launch?

11. What will it cost to run each month at our volume? A mid-sized business running one agent typically spends $200 to $1,500 a month, hosting included. A support agent handling 50,000 tickets a month spends $800 to $3,000 on model usage alone. A quote without a running-cost estimate is half a quote, and our AI agent cost guide shows how to work the number out.

12. What happens after launch, and if we part ways? You want the first month of fixes included, a named support option and a handover pack with documentation, architecture notes, the evaluation set and access to everything. Models change every few months, and someone has to rerun the tests when they do.

When build quotes arrive, compare them against these. They are our current AI development bands for US clients.

• AI feature added to an existing product: $10,000 to $25,000
• Focused custom AI app or chatbot on your own data: $15,000 to $40,000, 4 to 8 weeks
• AI agent with two or three system integrations: $18,000 to $40,000, 4 to 8 weeks
• Production RAG system: $18,000 to $35,000
• AI MVP for a startup: $30,000 to $70,000, 6 to 10 weeks
• Multi-tenant AI product: $45,000 to $150,000
• Support retainer: from $2,500 a month

Red flags that should end the conversation

If any of the following come up during those twelve questions, walk away.

• Demos only, with no production reference you can call
• A price before anyone has asked about your data or systems
• No evaluation set, or an accuracy figure promised before discovery
• Vague answers about which model is used or where it runs
• Accounts, keys or repositories held in the agency’s name
• An open-ended hourly contract for a clearly defined system
• “Guaranteed compliance” with laws they cannot name
• A chatbot or scripted automation sold as an “agent”, which Gartner calls agent washing

Your checklist before you sign

Run every shortlisted company through the same list and compare the answers side by side. Our case studies show the kind of shipped work you should ask every candidate to match.

• One production reference call completed
• Named engineers met and subcontractors disclosed
• Evaluation set size, pass rate and sign-off owner written into the scope
• A data flow diagram showing every model provider and its retention terms
• CCPA and ADMT position checked with your lawyer
• IP assignment clause signed, accounts in your company’s name
• Fixed price, fixed scope and a written change-request process
• Monthly running-cost estimate at your expected volume
• Handover pack and first month of fixes included
• If you are in the Bay Area, our San Jose office can walk through these questions with you on a free discovery call. Bring a competing quote and we’ll tell you where it’s fair

Frequently asked questions

Related reading